{"id":"e341c60e-ad27-4163-a5a5-c275aa1188c2","arxiv_id":"2412.05239","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A unified three-assumption theorem (contraction, local error, uniform control) converts finite-time bounds into global-in-time bounds for stochastic approximations.","lead":"This paper presents a general three-condition recipe for proving that an approximate random process stays close to the original process for all time, not just on finite intervals. The recipe is applied to averaging, numerical discretization, and mean-field particle systems, giving uniform-in-time error bounds.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Averaging example's averaged drift is miscomputed (missing r inside cosine), invalidating the claimed contraction range; theorem itself is sound.","rationale":"The central theorem and its proof are sound; the concern is a concrete error in the paper's advertised verification in the multiscale setting. The reader identified global contractivity as the weakest assumption, which is a scope limitation, but the most load-bearing issue is the miscomputed averaged drift in Example 2.10. This error invalidates the claimed verification for part of the parameter range and undermines the abstract's promise of examples where the joint conditions are verified. This does not alter the overall verdict of CONDITIONAL, because the framework itself remains valid; however, it strengthens the case that the examples need correction before the paper is used as a reference. The test is a direct, checkable computation of the invariant measure and the drift, so the concern can be settled quickly and objectively.","tokens_in":21893,"tokens_out":18268,"duration_ms":158913,"concrete_test":"Compute E[cos(Y)] for Y ~ N(r sin x, 1), obtaining e^{-1/2} cos(r sin x), and substitute into b_bar(x). Then rerun the synchronous-coupling contraction argument in Example 2.10 with this corrected drift: the Lipschitz constant of the drift becomes r, yielding the condition r <= e^{1/4}. If direct computation or simulation shows contraction actually holds up to r = e^{1/2} for the corrected drift, the concern is void; otherwise, the example must be corrected.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The general framework in Theorem 1.1 is internally correct: the telescoping-sum argument with a geometrically decaying contraction factor gives the uniform-in-time bound under the three assumptions. The load-bearing weakness is in the advertised verification, Example 2.10. For the slow-fast system (15)-(16), the frozen fast process at fixed x is dY = (-Y + r sin x)dt + sqrt(2)dB, whose invariant measure is N(r sin x, 1). Hence the averaged drift should be b_bar(x) = -x - r e^{-1/2} cos(r sin x). Equation (17) and the subsequent Itô calculation instead use cos(sin x), dropping one factor of r inside the cosine. This changes the Lipschitz constant of the drift: |d/dx cos(r sin x)| <= r, so the synchronous-coupling contraction calculation yields a rate -2(1 - r^2 e^{-1/2}), requiring r <= e^{1/4} ~ 1.284, not the claimed r <= e^{1/2} ~ 1.648. The verification of Assumption 1(1) and Assumption 2(1) in Example 2.10 is therefore invalid for part of the stated parameter range. Since the abstract promises 'examples where the joint conditions are verified', the averaging example does not currently support that claim. The theorem itself is not threatened, but the paper's demonstration of the framework in the multiscale setting is.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes three general sufficient conditions under which the distance between a stochastic process and an approximating process can be bounded uniformly in time: exponential contraction of one of the two semigroups in a chosen metric, a finite-time local error estimate, and a uniform-in-time control of a moment-type functional. The authors prove a general theorem (Theorem 1.1) using a telescoping-sum argument, then specialize the framework to averaging for slow-fast SDEs (Theorem 2.1 and Theorem 2.8), numerical discretisations of SDEs (Theorem 3.1), and mean-field particle systems (Theorem 4.1). Each section contains examples intended to verify the assumptions. The paper also includes a warning example, Appendix A, showing that time-inhomogeneous approximations need not admit uniform-in-time bounds. The central proofs appear sound; the main problems lie in the advertised worked examples, particularly the averaging example in Section 2.2.","tokens_in":22222,"tokens_out":17698,"duration_ms":155299,"significance":"If corrected, the framework would provide a useful unified rubric for a type of result that is usually proved case by case. The proof of Theorem 1.1 is transparent, does not rely on hidden parameter fitting, and applies to a general metric, which is a genuine strength. The paper is also honest about limitations: it flags the time-homogeneity requirement and explains why the roles of the two processes cannot be swapped in the averaging setting. However, the current version contains a material error in the averaging example: the averaged drift is miscomputed, which invalidates the verification of the contraction assumptions for the stated parameter ranges. A second example in the mean-field section states a rate that diverges as N grows. These issues affect the paper's claim to provide verified examples in all three advertised settings and require correction before the paper can be accepted.","major_comments":[{"comment":"The averaged drift is computed incorrectly. For the fast dynamics (16) frozen at x, the invariant measure is N(r sin x, 1), so E[cos(Y)] = e^{-1/2} cos(r sin x). Hence the averaged drift should be -x - r e^{-1/2} cos(r sin x), not -x - r e^{-1/2} cos(sin x) as printed in (17). This changes the Lipschitz constant of the drift from r e^{-1/2} to r^2 e^{-1/2}; the synchronous-coupling calculation in the example then yields contraction only for |r| ≤ e^{1/4}, not |r| ≤ e^{1/2}. The same incorrect drift also undermines the claim that Assumption 2(1) holds for |r| ≤ 3.5 because the drift is monotone: the derivative of the printed drift is -1 + r e^{-1/2} sin(sin x) cos x, and of the corrected drift is -1 + r^2 e^{-1/2} sin(r sin x) cos x, so the drift is not monotone for |r| up to 3.5 under the stated derivative criterion. The verification of Assumption 1(1) and Assumption 2(1) in this example is therefore invalid as written, and the example does not currently support the abstract's claim that the joint conditions are verified in the averaging setting.","section":"Section 2.2, Eq. (17) and Example 2.10"},{"comment":"The displayed propagation of chaos bound sup_{t≥0} E[N^{-1} ∑_{i=1}^N |\\bar X^i_t - X^{i,N}_t|] ≤ \\tilde C \\sqrt{N} cannot be correct as a convergence statement, because the right-hand side diverges as N → ∞. The intended bound is presumably \\tilde C / \\sqrt{N} or an equivalent N^{-1/2} rate. As printed, the example states a false conclusion.","section":"Section 4.2, Example 4.3"}],"minor_comments":[{"comment":"In the computation of d|Y^δ_t|^2, the martingale term is written with dW_t, but the fast process is driven by the independent Brownian motion B_t; this should read dB_t.","section":"Section 2.2, Example 2.10, Itô computation for Y^δ"},{"comment":"In the bound for |A_{1,2}|, the argument of φ is written as φ(\\bar X_τ, Y^δ_τ); it should be φ(X^δ_τ, Y^δ_τ), since the local error is evaluated at the state of the coupled process after one step.","section":"Section 2.2, proof of Theorem 2.8, k=2 step"},{"comment":"The final bound W2(νx \\bar p_t, (νp^δ_t)_x)^2 ≤ \\tilde C δ is equivalent to a δ^{1/2} bound in W2; the authors should state this explicitly to avoid confusion about the rate α in Assumption 1(2).","section":"Section 2.2, Example 2.10, final display"},{"comment":"The phrase 'verify first Assumption 4' appears to be a typo; presumably 'verify Assumption 4' is intended.","section":"Section 4.2, Example 4.3"}],"recommendation":"major_revision","confidential_remarks":"The core framework and the main theorems are sound, and the errors are localized to the worked examples. I therefore do not recommend rejection. However, the averaging example is the primary advertised verification in the multiscale section, and its current form contains a wrong averaged drift and incorrect contraction thresholds; the mean-field example also states a divergent rate. These should be corrected before the paper is reconsidered. The authors may also want to cross-check the constants in Example 2.10 against [11] and [5] carefully, since the verification now relies on imported results whose quantitative conclusions appear to have been applied too liberally."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe core framework is sound and worth knowing: the paper reduces uniform-in-time convergence to three checkable conditions—contractivity, local error, uniform control—and shows the same telescoping-sum argument works for averaging, numerical schemes, and mean-field systems. Theorems 1.1, 2.1, 3.1 and 4.1 are all the same geometric-sum idea, and the proofs are correct. The unified presentation is a legitimate contribution, even though the ingredients are known and the authors say so plainly, citing [51], [20], [11] and noting Theorem 2.8's proximity to [11, Theorem 3.2].\n\nThe soft spots are in the examples, not the framework. Example 2.10 miscomputes the averaged drift: the correct drift is -x - r e^{-1/2} cos(r sin x), but equation (17) drops the r inside the cosine. That changes the derivative by a factor of r, so the claimed contraction range r ≤ e^{1/2} and the monotonicity threshold |r| ≤ 3.5 are not supported. A synchronous-coupling bound for the corrected drift gives at best r ≤ e^{1/4} (or roughly 1.56 with a sharper estimate), not the stated ranges. The example needs to be redone. Example 4.3 also has a rate typo: it says uniform propagation of chaos gives E[N^{-1} Σ |X̄ - X^{i,N}|] ≤ C̃ √N, which diverges; it should be C̃/√N. That is minor.\n\nThe reliance on global exponential contractivity is a real limitation, but the authors don't oversell it. The literature review is honest, and the self-citations are transparent. The main theorems are sound and the unified framework is usable. I'd send this to peer review with a request to fix Example 2.10 and the rate in Example 4.3. Once those are corrected, I'd cite it for the framework.\n\nBest,\n\n[Your name]","headline":"Sound unified framework for uniform-in-time convergence, but Example 2.10 has a wrong averaged drift that breaks the advertised parameter ranges.","tokens_in":22712,"tokens_out":10591,"would_cite":true,"duration_ms":86030,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60J60","60H35","65C30","82C31"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that three sufficient conditions—exponential contractivity, a finite-time local error estimate, and a uniform moment control—turn finite-time approximation bounds into uniform-in-time bounds, and applies this recipe to…","keywords":["uniform in time convergence","multiscale methods","averaging for SDEs","numerical discretisation","mean-field particle systems","strong error","propagation of chaos","stochastic differential equations"],"falsifier":"Use the paper's Appendix A example as a test: for $dX_t=-X_t\\,dt+dW_t$ and $dX^\\delta_t=(-X^\\delta_t+\\mathbf{1}_{[1/\\delta,1/\\delta+1]})\\,dt+dW_t$, the processes are close on every fixed finite interval and share the same equilibrium, yet the error equals $1$ at $t=1/\\delta+1$, so no uniform-in-time bound holds. Since the only obstruction is that the approximating kernel is time-inhomogeneous, the decisive check is whether a time-homogeneous kernel can reproduce the same error runaway while satisfying all three General assumptions 1; if such a kernel exists the theorem is false, and if not, time homogeneity is an essential fourth condition.","tokens_in":21725,"feed_emoji":"⏳","tokens_out":10882,"duration_ms":98404,"temperature":0.7,"pith_summary":"Uniform-in-time convergence is the property that an approximation error stays small forever, not just on a finite horizon, which matters for long-time simulation and for interchanging the limits of time and of the approximation parameter. This paper proposes three sufficient conditions that together imply such a bound for any pair of time-homogeneous Markov processes: one of the processes contracts exponentially in a chosen metric, the two processes are close up to a fixed short time with error of order $\\delta^\\alpha$, and a certain 'size' functional of the approximating law stays bounded uniformly in time. The main theorem strings local errors along a grid and discounts each by the exponential contraction, so the sum is a geometric series whose constant does not grow with the horizon. The same template is then specialised and verified in three settings: averaging for slow-fast SDEs, numerical discretisation of SDEs, and mean-field particle systems, producing uniform-in-time strong error bounds and uniform-in-time propagation of chaos. Because the conditions are metric-agnostic and only use local-in-time estimates, existing finite-time convergence results can often be upgraded to global-in-time ones by checking the two extra structural conditions.","feed_headline":"Three conditions turn local errors into global-in-time bounds","feed_subtitle":"The same contractivity-plus-local-error argument covers averaging, numerical schemes, and mean-field particle systems.","key_machinery":"The load-bearing object is the pair of transition semigroups $p_t$ and $p^\\delta_t$ together with a metric $\\mathrm{dist}$ on probability measures. The proof mechanism is the telescoping sum along a grid: split time into blocks of length $\\tau$ (or length 1 in the discretisation case), compare the two processes only at grid points, and write the total distance as a sum of distances between neighbouring paths. Each term is bounded by the finite-time local error, then discounted by the exponential contraction factor $e^{-\\lambda \\cdot \\text{remaining time}}$; the uniform control makes the $M$-factor bounded at every grid point. The three assumptions are exactly what make each step of this sum work, and the sum converges as a geometric series, so the final constant does not grow with time.","core_discovery":"The paper's central claim is Theorem 1.1: whenever General assumptions 1 hold, namely (1) contractivity $\\mathrm{dist}(\\nu p_t,\\eta p_t)\\le e^{-\\lambda t}\\mathrm{dist}(\\nu,\\eta)$ for some $\\lambda>0$, (2) a finite-time local error $\\sup_{t\\le\\tau}\\mathrm{dist}(\\nu p_t,\\nu p^\\delta_t)\\le\\delta^\\alpha M(\\nu)$, and (3) a uniform control $\\sup_{t\\ge0}M(\\nu p^\\delta_t)\\le C(\\nu)$, then for every $t\\ge0$ there is a constant $\\widetilde C$, independent of $t$ and $\\delta$, with $\\mathrm{dist}(\\nu p_t,\\nu p^\\delta_t)\\le\\widetilde C\\delta^\\alpha$. The proof telescopes the distance over a time grid of step $\\tau$ and uses the contraction to weight each local error by $e^{-\\lambda(\\text{remaining time})}$, which makes the total a geometric sum. The same mechanism, with the roles of the processes adapted to each setting, yields Theorem 2.1 for the method of averaging, Theorem 3.1 for numerical discretisations, and Theorem 4.1 for mean-field particle systems; in Wasserstein metrics, strong-error corollaries (Corollary 2.2 and Corollary 3.2) follow as well. The authors present this as a common rubric that unifies existing isolated uniform-in-time results and lets finite-time estimates be leveraged into global ones.","pith_inferences":["Since $\\delta^\\alpha$ can be replaced by any $g(\\delta)\\to0$, the same template should cover approximation families with non-polynomial local error, such as projection or spectral approximations, provided their uniform control and contraction conditions can be verified.","The paper's 'leverage existing finite-time results' point implies a practical workflow for new approximations: establish a local error bound, then check exponential contraction of either side and a uniform moment bound; the uniform-in-time result follows automatically. The paper demonstrates the workflow but does not present it as a checklist.","The Appendix A counterexample suggests that time-homogeneity is doing essential work: approximations whose error-inducing mechanism acts on a time scale that diverges as $\\delta\\to0$ can satisfy local checks and still fail globally. Quantifying how the uniform constant must depend on such a time-scale ratio would be a natural extension to nearly time-homogeneous approximations.","Because the proof only uses the triangle inequality and the metric axioms needed for a pseudometric, the same three assumptions could be specialised to total variation or relative-entropy metrics by plugging in the corresponding contraction and local-error estimates, potentially yielding uniform-in-time entropic propagation of chaos."],"forward_implications":["For slow-fast SDEs, Theorem 2.1 gives $\\sup_{t\\ge0}\\mathrm{dist}(\\nu_x\\bar p_t,(\\nu p^\\delta_t)_x)\\le\\widetilde C\\delta^\\alpha$, and Corollary 2.2 converts this into a uniform-in-time strong $L^2$ error bound under a finite-time strong error condition.","For numerical discretisations, Theorem 3.1 yields $\\sup_{l\\in\\mathbb N}\\mathrm{dist}(\\nu\\pi_\\delta^l,\\nu p_{\\delta l})\\le\\widetilde C\\delta^\\alpha$ whenever the scheme has a local error bound and uniform moments; the paper verifies this for the Euler-Maruyama scheme for overdamped Langevin dynamics and for higher-order splitting schemes such as UBU.","For mean-field systems, Theorem 4.1 gives uniform-in-time propagation of chaos, $\\sup_{t\\ge0}\\mathrm{dist}((\\nu\\bar p_t)^{\\otimes N},\\nu^{\\otimes N}p_t^N)\\le\\widetilde C N^{-\\alpha}$, under contractivity of the particle system, finite-time chaos, and a uniform control of the limit law.","Because the bound is uniform in time, the limits $\\delta\\to0$ and $t\\to\\infty$ commute for the associated semigroups and observables, so the invariant measure of the approximation converges to the invariant measure of the limiting process (Remark 2.5).","In the general theorem and the numerical setting the two processes can be relabelled, so contractivity or uniform control may be checked on whichever side is easier; the averaging setting is the exception because the two laws live on different state spaces (Remark 2.4)."],"supporting_citations":[{"why":"Supplies the uniform-in-time averaging result most similar to Theorem 2.8 and the locally-Lipschitz Poisson-equation setting used to verify the assumptions.","marker":"[11]"},{"why":"Provides the finite-time strong and weak averaging error estimates used to verify Assumption 1(2) and Assumption 2(2).","marker":"[5]"},{"why":"Provides the strong exponential stability and uniform-in-time numerical scheme framework whose proof strategy underpins Theorem 2.8 and Remark 3.4.","marker":"[51]"},{"why":"Supplies the standard local error bound for the Euler-Maruyama scheme used to verify Assumption 3(2) in Corollary 3.5.","marker":"[17]"},{"why":"Provides the finite-time propagation of chaos estimate for McKean-Vlasov diffusions used to verify Assumption 4(2).","marker":"[39]"},{"why":"Supplies the Wasserstein contraction estimate for strongly convex Fokker-Planck dynamics used to verify Assumption 3(1).","marker":"[53]"},{"why":"Provides the reflection-coupling L1 contraction result that lets the uniform bound extend beyond strong convexity in Remark 3.6.","marker":"[54]"}],"fun_headline_variants":["Contract, control, and compute: global error bounds for approximations","Three conditions give global-in-time error bounds for approximations","One theorem unifies averaging, discretization, and mean-field limits","From local to global: a universal convergence criterion for processes","Global-in-time bounds via contraction and local error control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that at least one of the two processes has exponential contraction in the chosen metric; without a factor $e^{-\\lambda t}$, the discounted sum of local errors need not converge, and the paper's own examples only establish this contraction under strong convexity or small-parameter conditions.","fun_headline_variants_meta":{"raw":{"variants":["Contract, control, and compute: global error bounds for approximations","Three conditions give global-in-time error bounds for approximations","One theorem unifies averaging, discretization, and mean-field limits","From local to global: a universal convergence criterion for processes","Global-in-time bounds via contraction and local error control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000706,"raw_usage":{"total_tokens":3195,"prompt_tokens":972,"completion_tokens":2223,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":2141}},"tokens_in":588,"tokens_out":2223,"duration_ms":166115,"temperature":1.0,"reasoning_tokens":2141,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:51:18.222770+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use the paper's Appendix A example as a test: for $dX_t=-X_t\\,dt+dW_t$ and $dX^\\delta_t=(-X^\\delta_t+\\mathbf{1}_{[1/\\delta,1/\\delta+1]})\\,dt+dW_t$, the processes are close on every fixed finite interval and share the same equilibrium, yet the error equals $1$ at $t=1/\\delta+1$, so no uniform-in-time bound holds. Since the only obstruction is that the approximating kernel is time-inhomogeneous, the decisive check is whether a time-homogeneous kernel can reproduce the same error runaway while satisfying all three General assumptions 1; if such a kernel exists the theorem is false, and if not, time homogeneity is an essential fourth condition.","supporting_citations":[{"cited_title":"Poisson equations with locally-lipschitz coefficients and uniform in time averaging for stochas tic differential equations via strong exponential stability,","cited_arxiv_id":null,"evidence_quote":"Supplies the uniform-in-time averaging result most similar to Theorem 2.8 and the locally-Lipschitz Poisson-equation setting used to verify the assumptions."},{"cited_title":"Strong and weak convergence in the averaging principle for SDEs with H\\\"older coefficients","cited_arxiv_id":"1907.09256","evidence_quote":"Provides the finite-time strong and weak averaging error estimates used to verify Assumption 1(2) and Assumption 2(2)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the standard local error bound for the Euler-Maruyama scheme used to verify Assumption 3(2) in Corollary 3.5."},{"cited_title":"Propagation of chaos for a class of non-line ar parabolic equations,","cited_arxiv_id":null,"evidence_quote":"Provides the finite-time propagation of chaos estimate for McKean-Vlasov diffusions used to verify Assumption 4(2)."},{"cited_title":"Convergence to equilibrium in W asserstein distance for Fokker-Planck equations,","cited_arxiv_id":null,"evidence_quote":"Supplies the Wasserstein contraction estimate for strongly convex Fokker-Planck dynamics used to verify Assumption 3(1)."},{"cited_title":"Reflection couplings and contraction rates for diffu sions,","cited_arxiv_id":null,"evidence_quote":"Provides the reflection-coupling L1 contraction result that lets the uniform bound extend beyond strong convexity in Remark 3.6."}],"review_version":1}